cowpathInstall

the agent playbook

Correct your agent once.Not every week.

cowpath keeps the work you repeat with your agent: the start, your corrections, the final version. It turns them into a playbook, and the next run starts from it, yours and your team's.

Read the docsSet up a team →
curl -fsSL cowpath.sh/install | sh

macOS · Linux · free while in beta · Homebrew, npx and the MCP connector →

Works with

  • Claude Code
  • Codex
  • claude.ai
  • ChatGPT

The problem

The good version lives in last week's chat.

Every week you correct the same things, and every new chat starts from zero. The team's prompt doc went stale in spring, and five people make five versions of the same deck.

Claude

You · Mon 9:12

Again: no superlatives. Use their numbers, not ours.

Slack · #sales

Priya · 10:41

does anyone have the prompt dana used for the acme deck? it was in some chat in august

Notion

Sales prompts (v3 FINAL)

Last edited 214 days ago by Tom (deactivated)

ChatGPT

You · Thu 16:03

Like last time: Added, Fixed, Changed. One line each. No emoji in the headings.

Linear· SUP-412 · Support

Agent replies still promise refunds in 3–5 business days. Policy has been 10 since June.

Email

Re: Weekly update

Could these go back to the old format? Wins, risks, asks. Three bullets each.

Slack · DM

Lena (week 1) · 14:20

quick q: how do you all get claude to write support replies like yours? mine sound like a press release

ChatGPT

Memory

Memory updated: Prefers concise answers.

How it works

Correct it once. The next run knows.

01

Work as usual.

Hooks in Claude Code and Codex, or the connector in claude.ai and ChatGPT, keep each recurring run: the start, your corrections, the final version.

02

The playbook writes itself.

The first run writes the first playbook, and every few runs, your own agent (Claude Code, Codex, or the chat app) rewrites it from what held and what you corrected.

03

The next run starts there.

When the same work comes up, the playbook is in your agent's context before it writes a word.

01/16
Two corrections in the first run. None in the next.

Demo: release notes, corrected twice, then right the next week

  1. user: Write the release notes for v2.4 from the merged PRs.
  2. agent: v2.4 is our most powerful release yet. We've added blazing-fast CSV export, squashed a pesky sign-in bug, and much more.
  3. user: No superlatives. Group under Added and Fixed, one line per change.
  4. agent: Added - CSV export for reports Fixed - Sign-in loop after password reset - Timezone shift in scheduled exports
  5. user: Link the PR numbers. Customers click those.
  6. agent: Added - CSV export for reports (#804) Fixed - Sign-in loop after password reset (#812) - Timezone shift in scheduled exports (#815)
  7. cowpath: run kept · eng/release-notes · 2 corrections · final version
  8. cowpath: playbook v1 written · eng/release-notes
  9. cowpath: Group under Added and Fixed. One line per change.
  10. cowpath: No superlatives. Say what changed.
  11. cowpath: Link every PR number.
  12. cowpath: next week · new chat
  13. cowpath: playbook in context · eng/release-notes · v1
  14. user: Release notes for v2.5.
  15. agent: Added - Saved filters on the reports page (#841) Fixed - Duplicate rows in CSV export (#838)
  16. cowpath: 0 corrections

Use cases

The work you do every week, done your way the first time.

sales/decks/discovery

Before
Five reps, five decks, and the product tour on slide 2 again.
After
Every rep's agent opens with the customer's numbers.

Playbook · v3

  1. +Open with their numbers from the call notes, not ours.
  2. +Product tour: three slides, after the problem.
  3. +End on one next step with a date.

For teams

One playbook. Everyone's runs improve it.

Share a process.

Every agent on the team starts from the same playbook, the new hire's on day one included. When someone leaves, what they knew stays.

Every version has a name.

Who changed it, through which agent, from which run. Anyone with access can edit. Revert is one click.

Copy or share.

Copy a process and it's yours to change; it remembers where it came from. Share it and everyone improves the same one.

sales/decks/discovery
  1. V5 · MARA · CLAUDE CODE · RUN 41Open with the customer's numbers, not ours.
  2. V4 · YOU · WEBProduct tour down to three slides.
  3. V3 · JONAS · CHATGPT · RUN 33End on one next step with a date.

Your data

The transcript stays home. Only the run leaves.

On your machine.

The full transcript. The run is cut locally, by the agent CLI you already use, on your own plan. We never receive the chat.

With us, once you sign in.

The run (start, corrections, final version, rating) and the playbooks, visible to your workspace only. In claude.ai and ChatGPT, only what the agent passes to cowpath. Local mode sends nothing.

Yours to take.

Export any playbook as Markdown, anytime. Delete a workspace and its data goes with it.

In our tests

~35ms

added to a prompt

Median of 20 runs on a MacBook, Node start included.

3 of 3tasks

kept without being asked

Four vague prompts, none naming cowpath. The agent skipped the one-off question.

1run

to the first playbook

The first finished run writes it. The second starts from it.

0transcripts

uploaded

Runs are cut on your machine. Only the run leaves.

Measured in our own tests, October 2026.

Iterations per run

eng/release-notes

Example
playbook v1

eng/release-notes: 6, 7, 5, 6, 6, 5, 7, 6, 4, 3, 3, 2, 3, 2, 2, 1, 2, 1, 1, 1, 2, 1, 1, 1

run 1run 24

Install

One command. Then cowpath init.

  • curl -fsSL cowpath.sh/install | sh
  • brew install cowpath-sh/tap/cowpath
  • npx cowpath

claude.ai and ChatGPT: add https://cowpath.sh/mcp as a connector →

FAQ

Isn't this just memory?
No. Memory in chat apps keeps facts about one person, and you can't see or share what it kept. cowpath keeps one playbook per process: what good looked like, written down, versioned, shared with your team.
Is this observability?
No. Observability shows what happened. cowpath changes what happens next: the next run starts from a better playbook.
Do I have to change how I work?
No. In Claude Code and Codex the hooks keep the run. In claude.ai and ChatGPT your agent opens and closes runs itself. One-off questions are skipped.
What leaves my machine?
The run: the start, your corrections, the final version, your rating. Never the transcript. In local mode, nothing.
What does it cost?
Free while in beta. Team pricing comes later. Local mode is free.
What if a playbook gets worse?
Every version is kept, with a name and a run on it. Edit it, or revert in one click. A rewrite changes at most a fifth of the playbook.